Advance tracking of lateral objects

Abstract

The invention relates to a method for tracking a neighbouring vehicle in an environment of a main vehicle implemented by a driver assistance device of the main vehicle, which is connected to a camera capable of acquiring images within a field of view and at acquisition times, the method comprising: - detecting, from a first image acquired at a first acquisition time, a first lateral object with respect to the main vehicle; - determining an item of data associated with the first detected object; - creating a virtual object associated with parameters comprising a current position associated with the virtual object, wherein the current position is estimated from the item of data; - estimating, from the current position, a future position of the virtual object; - updating the parameters from the future position with, if a first criterion is met, an update of a current length of the virtual object.

Description

Description Title: Anticipatory tracking of lateral objects Technical field

[0001] This disclosure relates to the field of driver assistance systems and more particularly to the automatic detection of vehicles in areas of interest, in particular lateral areas. Prior art

[0002] The rise of Intelligent Transportation Systems (or ITS) has led to the development of numerous embedded systems in vehicles, particularly road transport. Such embedded systems include driver assistance systems and autonomous driving systems. In particular, the detection and / or tracking of objects, including vehicles, plays an important role in issues such as traffic flow, road safety, and road infrastructure management (e.g., variable message signs or speed cameras).

[0003] In the context of systems embedded in autonomous or semi-autonomous vehicles traveling in road traffic (e.g., on a motorway), early detection and tracking of objects surrounding the vehicle in question are particularly important in order to avoid collisions between the vehicle in question and other vehicles, for example. In particular, in lane change assistance or automatic lane change systems, early detection and reliable tracking of obstacles such as a central reservation or another vehicle located in a lane adjacent to the vehicle in question is required. Indeed, if the autonomous vehicle in question does not detect (or does not detect in an adequately anticipated or stable manner) another vehicle located in a neighboring lane and moves out of the way, a collision may occur between the two vehicles.

[0004] In such a context, most existing vehicle detection and tracking methods are based on classification algorithms for identifying vehicles from image sequences or video streams acquired by vision sensors (typically, a camera). Such vision sensors are generally arranged at the front and / or rear of the vehicle concerned, so as to acquire images and / or image sequences of a front and / or rear field of view (Field of View) of the vehicle concerned. The classification algorithms can then identify one or more vehicles neighboring the vehicle concerned on the basis of frontal recognition of the neighboring vehicles (i.e. by identifying and classifying vehicles detected in the field of view from their front or rear face).

[0005] However, such detection and tracking methods using vehicle classifiers prove ineffective when a front or rear face of neighboring vehicles is not yet detectable in the field of view, even if such neighboring vehicles are partially visible in the field of view, for example via a partial lateral face (or a side). Thus, if the vehicle concerned is located on a motorway lane and a neighboring vehicle is located relatively level with (or slightly behind) the vehicle concerned on a neighboring lane, a camera placed at the rear of the vehicle concerned would capture a rear lateral portion of the neighboring vehicle.In such a situation, the front of the neighbouring vehicle is not yet visible to the camera placed at the rear of the vehicle concerned, for example until the speed differential between the vehicle concerned and the neighbouring vehicle is large enough for the vehicle concerned to completely overtake the neighbouring vehicle. Until such complete overtaking occurs, the neighbouring vehicle is therefore substantially at the same level as the vehicle concerned and a lane change of the vehicle concerned into the neighbouring lane would result in a collision with the neighbouring vehicle. Such a criticality situation is then not detectable by existing classification methods.

[0006] Furthermore, most existing vehicle tracking methods rely on predefined dimensions associated with detected vehicle types. Such dimensions are generally not updated, so that even if vehicle tracking is implemented (e.g., by a classifier), such tracking does not allow reliable knowledge of the vehicle dimensions and therefore accurate tracking of the extent of the space occupied by the detected vehicle in the environment of the main vehicle.

[0007] There is therefore a need to secure the decision-making of driver assistance and / or autonomous driving systems, particularly in such a lane change decision context. In particular, there is a need for early and reliable detection, tracking and knowledge of vehicles surrounding a vehicle in question even before a front or rear face of such surrounding vehicles is visible in a field of view of a vision sensor of the vehicle in question - and therefore that conventional tracking of such surrounding vehicles by existing classifiers is possible. Summary

[0008] This disclosure improves such a situation.

[0009] A method is provided for tracking at least one neighboring vehicle present in an environment of a main vehicle, said neighboring vehicle and said main vehicle being motor vehicles, the method being implemented by a device configured to provide a driving assistance function for the main vehicle, said device being connected to at least one camera on board the main vehicle and capable of acquiring images of a scene surrounding the main vehicle according to at least one field of view and at acquisition times, the method comprising the following steps: - detect, from at least a first image acquired at a first acquisition time, a first object having a vertical plane extending laterally relative to the main vehicle, - determine, according to a predefined coordinate system, at least one piece of data relating to a first position associated with the first detected object, - create a virtual object associated with a set of parameters comprising at least one current position associated with the virtual object, said current position being estimated from at least said data relating to the first position according to the predefined coordinate system, - estimate, from at least said current position of the virtual object, a future position of the virtual object, and - updating the parameters associated with the virtual object from at least said future position, said updating of the parameters associated with the virtual object comprising, if a first criterion is satisfied, an update of a current length of the virtual object.

[0010] Therefore, the proposed method allows the implementation of an early tracking of a neighboring vehicle to the main neighbor, namely as soon as such a neighboring vehicle enters, even partially, into the field of view of a camera on board the main vehicle. In other words, the proposed method makes it possible to track a neighboring vehicle to the main vehicle despite the absence of visibility of a front or rear frontal face of the neighboring vehicle on which most existing classifiers rely for vehicle tracking.

[0011] Furthermore, the proposed method allows for the implementation of vehicle length update during the tracking process. Thus, unlike most existing classifiers, the proposed tracking method does not rely solely on pre-existing vehicle dimensions, but allows for updating the vehicle dimensions.

[0012] A main vehicle driving assistance function refers to a functionality of a system embedded in the main vehicle that can assist, guide or even decide on guidance of the main vehicle. Such a driving assistance function can be implemented in the context of a semi-autonomous or autonomous vehicle, for example, with or without a driver. A driving assistance function may include assistance with driving, changing lanes, overtaking another vehicle, or even parking, for example.

[0013] A field of view refers to a portion of the main vehicle's surroundings covered by the camera's field of view. In the context of a vehicle length update, the field of view considered is a rear field of view of the main vehicle.

[0014] By images acquired at acquisition times, we mean a continuous or discretized succession of images (i.e., in two dimensions) of the portion of the environment included in the field of view, associated with a succession of instants at which these images were acquired. In particular, the movement of the elements surrounding the main vehicle over time implies a field of view covering an evolving environment: the vehicles and elements neighboring the main vehicle exhibit a relative movement with respect to the main vehicle (as the main vehicle moves) and / or an absolute movement in the environment.

[0015] By a first object, reference is made to an object detected in the field of view. In particular, such an object may be associated with the movement (at least relative to the main vehicle) of a mobile body (i.e. non-inert) in the environment of the neighboring vehicle. Such a mobile body may for example be another vehicle, neighboring the main vehicle, an obstacle present in the environment or even a bird for example. Such a first object may for example correspond to an optical flow detected on the image, to a set of coplanar optical flows or segment detected on the image, to a polygon or an instance detected on the image (for example by a classifier). In particular, by a first object having a vertical plane extending laterally with respect to the main vehicle, reference is made to a detected object said to be lateral with respect to the main vehicle. In other words, in the field of view, a so-called lateral or oblique view of the object is detected.Such a side face is for example detected in the case of a neighboring vehicle traveling on a lane adjacent to the main vehicle and being downstream of the main vehicle.

[0016] By data relating to a first position associated with the first detected object, reference is made to an estimated or determined measurement and / or value making it possible to position the detected object in the acquired image and / or in the environment. Such data may, for example, correspond to a distance between the detected object and the main vehicle or to coordinates in two and / or three dimensions. The first position may, for example, refer to the position (in the environment and / or in the image) of a point belonging to the detected object.

[0017] A virtual object refers to a virtually modeled or created object that can be superimposed on an image acquired by the camera. Such a created virtual object is further associated with the detected object in the sense that the virtual object is estimated to be superimposed on a location presumed to be occupied by the detected neighboring vehicle. Such a virtual object is notably characterized by a position in the environment and / or in the image. Such a virtual object may also be characterized by dimensions so as to reflect presumed dimensions of the detected vehicle. Thus, the virtual object is created so as to model an occupation space of the neighboring body (e.g., a neighboring vehicle) detected via the detected object.

[0018] By a set of parameters, reference is made to parameter values ​​associated with a virtual object and therefore to an occupation space of the virtual object. Such a set of parameters may then include a current position of the virtual object, a current speed of the virtual object and / or current dimensions associated with the virtual object. Such a set of parameters then makes it possible to model the virtual object in a frame of reference of the image and / or the environment. In particular, the current position associated with the virtual object may refer to the position of a reference point associated with the virtual object, such a reference point being for example modeled as belonging to the front face (and therefore to the head) of the detected neighboring vehicle. Such a current position may in particular be imprecise in that the front face of the detected neighboring vehicle may not be visible in the field of view at the stage of implementation of the proposed method.Thus, the current position can be determined based on one or more hypotheses fixing such a current position in the image, for example on a border of the image.

[0019] By a future position, reference is made to a position estimated or predicted by a data estimator. In particular, the estimation or prediction of such a future position may be implemented based at least on current data fed to the estimator. In one embodiment, the future position of the virtual object is predicted and / or estimated by a Kalman filter from at least the current position of the virtual object.

[0020] According to another aspect, a device is proposed configured to provide a driving assistance function for a main vehicle, said device being connected to at least one camera on board the main vehicle and capable of acquiring images of a scene surrounding the main vehicle according to at least one field of view and at acquisition times, in which the device comprises at least one processing circuit configured to implement a method for tracking at least one neighboring vehicle present in an environment of the main vehicle.

[0021] According to another aspect, there is provided a computer program comprising instructions for implementing the tracking method when this program is executed by a processor.

[0022] According to another aspect, there is provided a non-transitory recording medium storing instructions for implementing the tracking method when executed, via a program, by at least one processor.

[0023] The features set out in the following paragraphs may, optionally, be implemented, independently of each other or in combination with each other:

[0024] In one embodiment, updating the current length of the virtual object includes extending the length of the virtual object by an amount determined from a difference between the current position and the future position, while keeping the current position of the virtual object constant.

[0025] Therefore, updating the parameters makes it possible to report the position difference between a current position and a predicted future position from at least one such current position over the length of the virtual object. In particular, the current position is maintained. Such a tracking method is then particularly advantageous, in particular when a front face of the detected neighboring vehicle is not yet visible in the field of view. Indeed, when a current position of the virtual object is, for example, associated with the position of a front face of the virtual object (modeling, for example, the front face of the detected neighboring vehicle), maintaining the current position makes it possible to faithfully model situations in which the neighboring vehicle is not yet completely visible (i.e., including the front face) in the field of view.Furthermore, once the front face of the vehicle is visible and detectable, for example by a classifier, the proposed method advantageously makes it possible to ensure continuity of tracking of the neighboring vehicle, with little variation between the location of the virtual object detected by the classifier and the location of the virtual object detected via the proposed method. Furthermore, the reporting of such a difference in position then makes it possible to estimate a length of the detected neighboring vehicle and therefore to update the space occupied by the virtual object already created. Thus, the proposed tracking method makes it possible to update tracking parameters of a vehicle different from the future position as estimated for example by data estimators.

[0026] In one embodiment, updating the current length of the virtual object is implemented if a first comparison between the future position and the current position indicates a moved back position of the virtual object in the first acquired image relative to the current position.

[0027] Therefore, the proposed tracking method allows updating the length of the virtual object modeling the neighboring vehicle based on a first comparison of the current and predicted positions of the virtual object. In particular, when it is detected that the virtual object is moving backward in the image (e.g., resulting in a lateral component of the future position being smaller than a lateral component of the current position), a systematic lengthening of the virtual object can be implemented. In particular, such an update of a length of the virtual object can rely only on current parameters and future parameters associated with the virtual object, without requiring new object detection on the acquired images.

[0028] In one embodiment, said data relating to the first position comprises a first lateral distance associated with the first detected object for the first acquisition time, said first lateral distance being determined from measurements of the first detected object in the first acquired image.

[0029] Therefore, the proposed tracking method makes it possible to determine a lateral gap existing between the detected neighboring body, for example, corresponding to the neighboring vehicle and the main vehicle. The first position corresponding, for example, to the position of a specific point of the first detected object can in particular make it possible to deduce information concerning the proximity of the detected neighboring vehicle in the environment relative to the main vehicle.

[0030] By a first lateral distance, we mean a lateral or bias gap between the element associated with the first detected object and the main vehicle. For example, in the case of a main vehicle and a neighboring vehicle traveling on two parallel traffic lanes (e.g., two highway lanes), the lateral distance corresponds to the lateral gap separating the lanes occupied by the main and neighboring vehicles. Such a lateral distance thus reflects a gap in the environment between the main and neighboring vehicles.

[0031] In one embodiment, the future position is associated with a lateral object distance of the virtual object.

[0032] Therefore, once the virtual object is created, its displacement in the environment can be estimated, for example by a data estimator, so that predicted, or future, parameters such as a future position and an object lateral distance can be estimated for such a virtual object.

[0033] In one embodiment, the tracking method further comprises: - obtain, from at least a second image acquired at a second acquisition time subsequent to the first acquisition time, data relating to a second position associated with a second detected object, said data relating to the second position comprising a second lateral distance associated with the second detected object for the second acquisition time, - if the second detected object is determined to belong to the virtual object, comparing, in a second comparison, the second lateral distance and the object lateral distance, wherein updating the parameters associated with the virtual object further depends on a result of the second comparison.

[0034] Consequently, the proposed tracking method makes it possible to take into account a change in the occupancy of the neighboring vehicle detected in the environment. Thus, on a second image acquired at an acquisition time subsequent to the first acquisition time, the second detected object potentially differs from the first detected object and makes it possible to update the change in the space occupied by the neighboring vehicle in the environment (for example, the neighboring vehicle “enters more and more” into the field of view, or the speed differential between the main and neighboring vehicles is such that the main vehicle has completely overtaken the neighboring vehicle and a front face of the latter is now visible in the field of view). The proposed update therefore makes it possible, based on a second criterion distinct from the first criterion, to update the virtual object.

[0035] A second detected object refers to a distinct object, disjointed or not from the first detected object on an acquired image. Such a second detected object can then correspond to an instance, a polygon, an optical flow, a segment reflecting a movement of the neighboring vehicle detected in the environment. The second position associated with such a second detected object can then reflect a new space occupied by the neighboring vehicle.

[0036] In one embodiment, if the second lateral distance is strictly less than the object lateral distance, the update of the parameters associated with the virtual object comprises an update of the current position of the virtual object, said update of the current position including at least one update of a lateral component of the current position of the virtual object corresponding to a repositioning of the virtual object.

[0037] Therefore, the length update is conditioned by a comparison of the lateral distances from the successively detected objects. In particular, if the second lateral distance is greater than or equal to the estimated lateral distance of the virtual object, an update of the current position of the virtual object is not required since it is considered that the neighboring vehicle gradually enters the field of view. Thus, the proposed tracking method takes advantage of the potential absence of a visible frontal face in the field of view to maintain the current position of the vehicle (and thus avoid model a virtual object “receding” in the acquired image) and compensate for the difference between the current and future positions by a lateral extension of the length of the neighboring vehicle.

[0038] In particular, the method advantageously makes it possible to correct potential overestimations of the determined and / or predicted lateral distances between the neighboring vehicle and the main vehicle. Such a correction thus allows the tracking method to gain in precision, and to avoid dangerous situations, in which the neighboring vehicle would in fact be closer to the neighboring vehicle than estimated (second lateral distance corresponding to the second object detected at a more recent time shorter than the estimated object lateral distance).

[0039] Therefore, the proposed tracking method allows to take into account inaccuracies related to the creation of the virtual object based on the parameters and the current position. Indeed, if the second lateral distance is strictly less than the object lateral distance, the virtual object estimated on the basis of the object lateral distance does not reflect the space actually occupied by the neighboring vehicle and in particular, the proximity of the neighboring vehicle to the main vehicle has been underestimated. The proposed method then allows to rectify such a potentially critical situation (since the neighboring vehicle is potentially closer to the main vehicle compared to the location modeled by the virtual object via the current position), by updating the lateral position of the virtual object so as to reflect the actual lateral proximity between the neighboring and main vehicles.Thus, the method advantageously makes it possible to maintain consistency between the updating of the virtual object and what is detectable in the second image.

[0040] In one embodiment, if the second lateral distance is strictly less than the object lateral distance, updating the parameters associated with the virtual object comprises updating the current position of the virtual object, said updating of the current position further including updating a height component of the current position of the virtual object corresponding to a repositioning of the virtual object. Such an update of the height component may further be based on terrain topology information in the environment.

[0041] In one embodiment, the current position of the virtual object is associated with the position of a first reference point belonging to the virtual object, the position of the first reference point belonging to the virtual object being determined such that, in the predefined coordinate system corresponding to a reference image in the acquired images, the first reference point is positioned on a vertical end border of the first image.

[0042] Consequently, the current position makes it possible to model the location of the neighboring vehicle in the environment despite a partial side view of the vehicle thus detected. The current position then obeys constraints by modeling the location of the neighboring vehicle on the vertical border of the image.

[0043] By a first reference point, reference is made to a point estimated to belong to the neighboring vehicle detected in the environment. In particular, such a first reference point may model the location of a front face of the neighboring vehicle (which is potentially not yet visible in the field of view).

[0044] In one embodiment, the virtual object is created from the position of the virtual object and predefined dimensions associated with a type of motor vehicle, at least one predefined dimension corresponding to the length of the virtual object.

[0045] Therefore, the proposed tracking method makes it possible to model, via the virtual object, a space occupied by the neighboring vehicle, despite the absence (at least during an initial iteration of the tracking method) of information relating to the dimensions of the detected vehicle.

[0046] By the length of the virtual object, reference is made to a lateral length of the virtual object, such virtual object modeling a neighboring vehicle from an object detected obliquely (along a lateral plane) relative to the main vehicle.

[0047] In one embodiment, the future position of the virtual object is estimated from a plurality of positions of the virtual object respectively determined for the acquisition times until detection of at least one frontal portion of the neighboring vehicle in the field of view of the camera on at least one of the acquired images.

[0048] Therefore, the proposed tracking method can advantageously be implemented before a front (or frontal) face of the neighboring vehicle is visible, so as to allow early tracking of neighboring vehicles to the main vehicle. Thus, when a front (or frontal portion) of the neighboring vehicle becomes visible in the field of view, existing classifiers can take over vehicle tracking by updating the virtual object.

[0049] In one embodiment, each of the plurality of positions of the virtual object is associated with an uncertainty value relating to an uncertainty in the initial conditions, said uncertainty value being greater than a predefined uncertainty threshold.

[0050] Therefore, the parameters determined for modeling the virtual object based on a biased view of the detected neighboring vehicle can be assigned an uncertainty value larger than a threshold value, corresponding for example to an uncertainty value of data determined by a conventional classifier. Thus, when the parameters determined in the proposed method are fed to a data estimator for predicting a future position, a weight relative to a degree of error is assigned to the parameters. Such an uncertainty value may, for example, correspond to a variance or a covariance. Thus, when the estimator predicting the future position receives other values ​​of the parameters with a lower degree of uncertainty, the estimator can predict the future position of the virtual object on the basis of these values. For example, when a front face of the neighboring vehicle becomes visible in the field of view, a classifier of vehicle front faces can then provide the estimator with more reliable data than data from a position estimation from a detected lateral object.The proposed method then makes it possible to perform a transition between different data sources allowing the tracking of the neighboring vehicle.

[0051] In one embodiment, the parameters associated with the virtual object are updated by keeping a parameter of the virtual object constant.

[0052] Therefore, the proposed method makes it possible to maintain stability of the virtual object when updating the parameters ensuring the tracking of the neighboring vehicle. Indeed, the proposed method makes it possible to update both a length and / or a position of the virtual object. A parameter associated with the virtual object is then kept constant, for example at least one component of the position associated with the virtual object, when the length of the virtual object is updated, or a temporal parameter associated with the virtual object such as a collision time or the product of the speed by a distance of the virtual object, when the position of the virtual object is updated. Brief description of the drawings

[0053] Other features, details and advantages will become apparent upon reading the detailed description below, and upon analyzing the attached drawings, in which:

[0054] [Fig. 1] Figure 1 shows a schematic of a main vehicle according to one embodiment.

[0055] [Fig. 2] Figure 2 shows an aerial view of a scene surrounding the main vehicle according to one embodiment.

[0056] [Fig. 3] Figure 3 shows a diagram of a driving assistance device according to one embodiment.

[0057] [Fig. 4] Figure 4 shows steps of a method for tracking a vehicle according to one embodiment.

[0058] [Fig. 5] Figure 5 shows a shot of a scene surrounding the main vehicle at a first acquisition time according to one embodiment.

[0059] [Fig. 6] Figure 6 shows a virtual tracking object created according to one embodiment.

[0060] [Fig. 7] Figure 7 shows an update of the virtual tracking object created according to one embodiment.

[0061] [Fig. 8] Figure 8 shows a shot of a scene surrounding the main vehicle at a second acquisition time according to one embodiment.

[0062] [Fig. 9] Figure 9 shows an update of the virtual tracking object created according to one embodiment.

[0063] [Fig. 10] Figure 10 shows a shot of a scene surrounding the main vehicle at a third acquisition time according to one embodiment.

[0064] [Fig. 11] Figure 11 shows a measurement carried out on a shot of a scene surrounding the main vehicle at a first acquisition time according to one embodiment.

[0065] [Fig. 12] Figure 12 shows a reference point associated with a neighboring vehicle according to one embodiment. Description of the embodiments

[0066] Reference is made to Figure 1. Figure 1 shows a diagram of a main passenger vehicle (PC). The main passenger vehicle (PC) may be a motor vehicle. There is no limitation on the type of vehicle to which the main passenger vehicle (PC) belongs. The main passenger vehicle (PC) may, for example, be a private, utility, or industrial vehicle, and may, for example, correspond to a car, a van, a two-wheeler, a truck, or a bus. The main passenger vehicle may also be a towed vehicle, for example, towing a trailer, a semi-trailer, or a caravan. The dimensions of the main passenger vehicle may be between 2 meters and 20 meters in length, between 0.5 meters and 5 meters in width, and between 1 meter and 5 meters in height.

[0067] The main vehicle (VP) is equipped with at least one on-board system providing a plurality of functions or applications of the main vehicle (VP). Such functions may, for example, correspond to cruise control, power steering, automated airbag deployment, automatic headlight adjustment, etc.

[0068] The main vehicle VP is in particular equipped with an on-board system allowing detection of objects surrounding the main vehicle VP, for example as part of a driving assistance function such as obstacle detection, assistance with Lane change or automatic lane change. For this purpose, the on-board system of the main vehicle VP includes a driving assistance device 2. The device 2 is capable of providing a driving assistance function based on a plurality of views (or images) of the environment of the main vehicle VP. For this purpose, the device 2 is in particular connected to one or more vision sensors such as a camera device or a camera 1 (the vision sensor will be considered to be a camera 1 in the following description). The device 2 can also be capable of transmitting or communicating data, for example related to the driving assistance function provided. For this purpose, the device 2 can be connected to a communication interface 50. In a particular embodiment, the interface 50 can be integrated into the device 2. Such a communication interface 50 can for example be a human-machine interface.The interface 50 may include a display screen, a touch screen, a dashboard, and / or a speaker. The data transmitted by the device 2 to the interface 50 may, for example, correspond to assistance information indicating to the driver of the main vehicle VP whether or not he can change lanes.

[0069] Reference is now made to Figure 2. Figure 2 illustrates an ENV environment (or scene), in which the main vehicle VP is located. Figure 2 is an aerial view of such an ENV scene.

[0070] The ENV environment can be defined in a three-dimensional frame (X,Y,Z), called a "world frame" as illustrated in Figures 1 and 2. The origin of such a world frame (X,Y,Z) is predefined and fixed in the ENV environment.

[0071] The main vehicle VP is considered to be moving in the ENV scene. Such an ENV scene corresponds, for example, to a road or a motorway consisting of several traffic lanes. These traffic lanes may in particular be parallel to each other, as represented by the linked dotted vertical lines in Figure 2. Figure 2 illustrates, for example, three traffic lanes, with the main vehicle VP being located on the middle traffic lane. The main vehicle VP is considered to be moving in the main direction X of the (X,Y,Z) frame, such a main direction being called the longitudinal direction. Such a movement is shown diagrammatically in Figures 1 and 2 by an arrow attached to the main vehicle VP. The main vehicle VP may also have a movement in a Y direction of the (X,Y,Z) frame, called the lateral direction.Such a movement in the Y direction, called lateral movement (or displacement), can for example take place when the main vehicle VP changes lane. A movement in the Z direction, i.e. in height, of the main vehicle VP is considered absent or negligible. The main vehicle VP is therefore considered to be kept on the ground and therefore has a constant height in the Z direction, corresponding to a predefined dimension. of the main vehicle VP. In one embodiment, such a height in the Z direction of the main vehicle VP may vary in the order of a centimeter or a decimeter, such a variation in height in the Z direction being for example linked to shock absorbers of the main vehicle VP and / or to reliefs or roughness present in the ENV environment (in particular on the traffic lane of the main vehicle VP). In the remainder of the description, such a height in the Z direction of the main vehicle VP is considered known.

[0072] The scene ENV surrounding the main vehicle VP also includes other OV elements and vehicles VV1, VV2, VV3. The vehicles VV1, VV2, VV3 are neighboring vehicles of the main vehicle VP. Such neighboring vehicles VV1, VV2, VV3 are, like the main vehicle VP, motor vehicles moving in the scene ENV. Such neighboring vehicles VV1, VV2, VV3 may in particular have dimensions and movement characteristics (in terms of speed or acceleration for example) similar to or different from the main vehicle VP. For example, figure 2 may represent a main vehicle VP moving on a motorway lane and neighboring vehicles VV1, VV2, VV3 traveling on the motorway lanes neighboring the lane taken by the main vehicle VP. OV objects are neighboring elements of the main vehicle VP and can refer to any distinct element of a neighboring vehicle VV1, VV2, VV3.A neighboring element OV can, for example, correspond to an obstacle located in the scene such as a central reservation separating two traffic lanes, an indication sign or even a bird flying in the ENV scene.

[0073] As illustrated in Figure 2, the main vehicle VP is considered to be equipped with a camera 1 positioned at the rear of the main vehicle VP. In another embodiment (not shown in Figure 2), the camera 1 may be positioned at the front of the main vehicle VP or several cameras 1 may be positioned both at the front and at the rear of the main vehicle VP. In the context of the present description, the camera 1 is considered to be positioned at the rear of the main vehicle VP and the field of view FOV is a rear field of view of the main vehicle VP. The camera 1 makes it possible to capture images (or shots) of the scene ENV of the main vehicle VP according to a field of view FOV (or in English “Field of View”). Such a field of view FOV depends in particular on the type of camera 1 used and the positioning of the camera 1 in (or on) the main vehicle VP.With reference to Figure 2, the field of view FOV covers a portion of the scene ENV, such that a limited portion of the scene ENV is captured according to the field of view FOV. Thus, in Figure 2, the field of view FOV represented covers a portion of the traffic lane in which the main vehicle VP is located and respective portions of the neighboring traffic lanes. In particular, the field of view FOV covers a portion of. the scene ENV in which the neighboring vehicle VV2 is located in its entirety and a portion of the scene ENV in which the neighboring vehicle VV1 is partially located (unhatched gray area). For example, as illustrated in Figure 2, a left rear end of the neighboring vehicle VV1 (for example including the left rear wheel of the neighboring vehicle VV2) belongs to the field of view FOV. However, a reference point W, for example located at the middle of the front face of the vehicle, is not in the field of view FOV in Figure 2. The neighboring vehicle VV3 is not visible in the field of view FOV of the main vehicle VP. A portion of the neighboring vehicle VV1 is not visible in the field of view FOV of the main vehicle VP. The non-visible parts of the neighboring vehicles VV1, VV3 in the field of view FOV are represented in Figure 2 by striped areas. A neighboring object OV, for example a bird, may be visible in the field of view FOV.

[0074] In the context of an ENV scene as represented in figure 2 according to the (X,Y,Z) reference frame, it is for example considered that the main vehicle VP moves at a main speed known. To facilitate the rest of the description, such a main speed can be considered to be of constant VVP standard. The neighboring vehicle VV1 is considered to be moving at a neighboring speed unknown to the driving assistance device 2 of the main vehicle VP. In the embodiment described below, it can be considered that the VVP standard of the neighboring speed is lower than the VVP standard of the main speed V^. Under such an assumption and under the assumption that the neighboring speed of the neighboring vehicle VV1 remains substantially constant over a considered time interval, like the main speed of the main vehicle VP, the distance difference in the longitudinal direction X between the main vehicle VP and the neighboring vehicle VV1 will increase over time, so that at a future time, the overtaking of the main vehicle VP on the neighboring vehicle VV1 will be such that the neighboring vehicle VV1 will be entirely included in the (rear) field of view FOV of the main vehicle VP, as is the initial case of the neighboring vehicle VV2.

[0075] Such an evolution of the membership of the neighboring vehicle VV1 to the field of view FOV of the main vehicle VP over time is illustrated in Figures 5, 8 and 10.

[0076] Reference is made to Figures 5, 8 and 10. Figures 5, 8 and 10 diagrammatically show shots of the scene ENV surrounding the main vehicle VP according to the field of view FOV of camera 1 of the main vehicle VP, as positioned in Figure 2. In order to facilitate the readability of Figures 5, 8 and 10, only the portion of the neighboring vehicle VV1 visible in the field of view FOV is shown in Figures 5, 8 and 10, the neighboring vehicle VV2 and the neighboring element OV of Figure 2 included in the field of view FOV are not shown in the shots of Figures 5, 8 and 10.

[0077] Figures 5, 8 and 10 may correspond to images successively acquired by the camera 1 at successive instants (or acquisition times) T1, T2 and T3. Each of the instants T1, T2 and T3 may for example be spaced one or more milliseconds apart in time. Each of the images corresponds to a set of pixels with definable coordinates in a two-dimensional frame (y, z), called the “image frame” as shown in Figures 5, 8 and 10. For example, the lowest and leftmost pixel of each image acquired by the camera 1 may correspond to the origin of the image frame (y, z). In other embodiments, the central pixel of the image acquired by the camera 1, the optical center of the image, the pixel of the image associated with the location of the camera 1 in the environment ENV or even the pixel corresponding to a vanishing point C' of the acquired image can be considered as the origin of the reference (y,z).The resolution of the acquired images (and therefore the pixels) is the same for all acquired images and depends in particular on the properties of the camera 1 .

[0078] Figure 5 shows a diagram of an image acquired by the camera 1 at time T1, corresponding for example to a configuration of the scene ENV and the field of view FOV illustrated in figure 2. Thus, only a first so-called lateral portion of the neighboring vehicle VV1 is visible in the image of figure 5, such a first lateral portion including in particular the left rear wheel of the neighboring vehicle VV1.

[0079] Figure 8 shows a diagram of an image acquired by the camera 1 at time T2 following time T1. At such a time T2, a second lateral portion of the neighboring vehicle VV1 larger than the first lateral portion is visible in the image of Figure 8, such a second lateral portion including in particular always the left rear wheel, and partially the left front wheel of the neighboring vehicle VV1. In other words, the speed differential between the main vehicle VP and the neighboring vehicle VV1 between times T1 and T2 is such that the neighboring vehicle VV1 “enters more and more” into the field of view FOV of the camera 1 between times T1 and T2.

[0080] Figure 10 shows a diagram of an image acquired by the camera 1 at time T3 following time T2. At such a time T3, the entire neighboring vehicle VV1 is visible in the image of Figure 10. In particular, a reference point W shown in Figure 2 and positioned on the front end of the neighboring vehicle VV1 has become visible in the acquired image, such a reference point W corresponding to the pixel w in the image of Figure 10. In other words, the speed differential between the main vehicle VP and the neighboring vehicle VV1 between times T2 and T3 is such that the neighboring vehicle VV1 “entered entirely” into the field of view FOV of the camera 1 between times T2 and T3.

[0081] In the context of a driving assistance function provided by the device 2, for example aimed at assisting the main vehicle VP in a lane change maneuver, a process for detecting and tracking vehicles surrounding the main vehicle VP is required, so that the main vehicle VP does not move onto a neighboring traffic lane if it risks colliding with one of the surrounding vehicles. Existing techniques for detecting and tracking vehicles surrounding the main vehicle VP are based in particular on the use of classifiers, based for example on convolutional neural network (CNN) methods, K-nearest neighbors (KNN) or support vector machines (SVM).Such classifiers allowing the detection and tracking of vehicles surrounding the main vehicle VP rely in particular on the detection and classification of a frontal (front or rear) view (or face) of the vehicles surrounding the main vehicle VP. The use of such classifiers may in particular comprise a learning phase based on a plurality of images representing frontal views of various types of vehicles. Thus, with reference to FIG. 2, a front frontal view of the neighboring vehicle VV2 belonging to the field of view FOV of the camera 1, the detection and tracking of the neighboring vehicle VV2 can be implemented on the basis of existing classification techniques. The same applies to the neighboring vehicle VV1 in the image of FIG. 10 at time T3.

[0082] However, the positions of the neighboring vehicle VV1 shown in Figure 2 and Figures 5 and 8 correspond to situations in which the existing classifiers fail to detect or track the neighboring vehicle VV1 effectively, or in any case, not without generating significant costs and / or computational time, due to the absence of a visible frontal view of the neighboring vehicle VV1 in the shots of Figures 5 and 8 (the reference point W of Figure 2 positioned on the front frontal face of the neighboring vehicle VV1 not yet being visible in the field of view FOV of camera 1 at these stages).These situations, however, correspond to critical situations during which the main vehicle VP could collide with the neighboring vehicle VV1 if the main vehicle VP changes position and moves into the lane of the neighboring vehicle VV1, in the absence of lane change assistance information indicating that the neighboring vehicle VV1 is detected near such a position.

[0083] A method for early tracking of the neighboring vehicle VV1 is then proposed and detailed in Figure 4 to detect and track the neighboring vehicle VV1, and in particular at the stage of the situations represented in Figures 2, 5 and 8. Such a stage is said to be “early” in that a front face of the tracked vehicle may not yet be visible in the field of view FOV. Such a method of early tracking of the neighboring vehicle VV1 is implemented by a driving assistance system shown in Figure 3.

[0084] Reference is now made to Figure 3. Figure 3 represents a diagram of an on-board system of a main vehicle VP. In particular, such an on-board system corresponds to a driving assistance system for the main vehicle VP. The driving assistance system of the main vehicle VP makes it possible in particular to provide a lane change assistance or automatic lane change function for the main vehicle VP, when the main vehicle VP is moving on a traffic lane as shown for example in Figure 2.

[0085] The system firstly comprises the driving assistance device 2. The device 2 itself comprises a unit 20 for detecting objects on acquired images, a unit 30 for determining parameters associated with the objects detected by the unit 20, a unit 40 for creating virtual objects associated with the detected objects, a unit 50 for estimating data, such as a Kalman filter, (for example by fusion, by learning and / or by prediction of data) for tracking created virtual objects and a unit 60 for updating parameters associated with the created virtual object.

[0086] The driving assistance device 2 is further connected, on the one hand, to a vision sensor of the camera or camera type 1. The device 2 may also comprise an input unit (not shown in FIG. 3) allowing the device 2 to receive in substantially real time a data stream from the camera 1. Such a data stream corresponds to a discrete or continuous succession of images (or shots) of the scene ENV according to the field of view FOV of the camera 1. Each image received is associated with an acquisition time of the image by the camera 1. Each image may be time-stamped.

[0087] Furthermore, the driving assistance device 2 may be connected to a communication interface 70. Such an interface 70 may correspond to a human-machine interface integrated into the on-board system of the main vehicle VP. Such an interface 70 may also be integrated into the device 2. The communication interface 70 may also be a remote interface. The communication interface 70 may comprise a display screen, a touch screen, a dashboard or even a loudspeaker, making it possible to transmit, for example by visual, haptic and / or audible information, indications relating to assistance in driving the main vehicle VP. In particular, the communication interface makes it possible to transmit a detection status of a vehicle neighboring the main vehicle VP or an indication relating to a possibility of changing lane of the main vehicle VP. Such information transmitted by the interface 70 may for example correspond to an estimated aerial visual representation of the respective positions of the main vehicle VP and the neighboring elements belonging to the field of view FOV in real time (eg, as represented in FIG. 2), an audio stimulus alerting of a risk of collision, or even a superposition of a virtual object OV1 highlighted and updated in real time on a stream of images coming from the camera 1, as illustrated for example in FIGS. 6, 7 and 9.

[0088] The units 20, 30, 40, 50 and 60 of the device 2 each comprise a processing circuit including at least one processor (21, 31, 41, 51, 61) and a memory unit (22, 32, 42, 52, 62) in order to implement one or more steps of the method for early tracking of a neighboring vehicle VV1, which will be described in FIG. 4. In particular, each processing unit 20, 30, 40, 50 and 60 of the device 2 can rely on data processed and / or obtained by other units in order to implement one or more steps of the method for early tracking of a neighboring vehicle VV1, as illustrated by the arrows in FIG. 3.

[0089] Reference is now made to Figure 4. Figure 4 illustrates a succession of steps for implementing a method for early tracking of a neighboring vehicle VV1 by a system including a device 2 for assisting in driving a main vehicle VP as shown in Figure 3. In particular, the method for early tracking of a neighboring vehicle VV1 described in Figure 4 comprises a phase of detecting an object associated with the neighboring vehicle VV1 and a phase, strictly speaking, of tracking the neighboring vehicle VV1 from the detected object.

[0090] In a step 400, the device 2 receives a plurality of images (or shots) associated with respective acquisition times of the images from the camera 1. Such images may in particular be received continuously, for example via a video stream. In such a case, in step 400, the device 2 may discretize the received video stream so as to obtain a set of discrete time-stamped images associated with respective acquisition times.

[0091] Steps 410 to 450 described below are implemented from an object detected on an acquired image - for example a first detected object OBJ1 - at a given acquisition time - for example the first acquisition time T1 -.

[0092] In a step 410, a first object OBJ1 is detected on at least one of the acquired images. With reference to FIG. 5, such a first detected object OBJ1 is represented on the image corresponding to the first acquisition time T1. In one embodiment, the first detected object OBJ1 is a so-called lateral object in the acquired image, in that the first detected object OBJ1 belongs to a lateral plane of the image, such a lateral plane being for example parallel to the plane (X,Z) in the world reference frame (X,Y,Z). The detection of such a first lateral object OBJ1 is then said to be anticipated in that it takes place before a front face of the neighboring vehicle VV1 is visible. In particular, in step 410, the first detected object OBJ1 corresponds to a non-inert body. In other words, the first detected object OBJ1 corresponds to a body moving in the environment ENV (and therefore, in the world frame (X,Y,Z)) and having a speed (for example, a so-called longitudinal speed corresponding to a movement along the X axis in the world frame (X,Y,Z) and / or a so-called lateral speed corresponding to a movement along the Y axis and / or a rotation speed).

[0093] The first object OBJ1 may be detected in step 410 on a first image (for example, the image of FIG. 5) by processing one or more successively acquired images. In one embodiment, the first object OBJ1 may for example be detected by a classifier capable of detecting a lateral object OBJ1 on the acquired image, for example a classifier detecting a lateral portion of a vehicle, such as a vehicle wheel detector. Alternatively, the first object OBJ1 may be detected by homographies on at least two successively acquired images, so as to identify one or more related sets, or optical flows, of relatively homogeneous speed on the images in a delimited time interval. According to such a variant, the detection of the first object OBJ1 then includes determining a displacement of pixels at a substantially common speed from one image to the next, or optical flow.Each optical flow can be determined by matching points between several successive images, for example by the Lucas-Kanade method. The determined optical flows can in particular be segmented into one or more segments having a common point or vanishing line and belonging to lateral planes in the field of view FOV. In other words, the acquired image can be segmented into one or more segments, each segment defining a grouping of optical flows (and therefore a grouping of pixels having substantially the same speed and describing the same movement in a given time interval).

[0094] Thus, in step 410, the first detected object OBJ1 is a set of pixels that can be formed by an instance, an outline, a polygon or even a rectangle (or in English, a "bounding box") detected and classified by a classifier by segmentation of objects in the acquired image for example. Alternatively, the first detected object OBJ1 is a set of pixels that can be formed by an optical flow or segment including several optical flows. In particular, the set of pixels forming the first detected object OBJ1 can be a set of coplanar pixels parallel to a lateral plane in the world reference frame (X,Y,Z).

[0095] In a step 420, data relating to a first position of the first detected object OBJ1 can be determined by the device 2. In particular, such data relative to the first position of the first object OBJ1 can be a first lateral distance di at,i associated with the first detected object OBJ1. Such a first lateral distance diat,i may in particular be determined from values, measurements, estimates or data associated with the first detected object OBJ1 on the first acquired image. In one embodiment, such a first lateral distance di a t,i may for example be a value fixed approximately taking into account an area of ​​the acquired image in which the first detected object OBJ1 is located. In one embodiment, such a first lateral distance di at ,i can be determined from a number of ground lines visible on the first acquired image between a central vertical line VL of the image (for example, the dotted vertical line) and a ground line at which the first detected object OBJ1 is located, for example in the case of vehicles on motorway lanes. In one embodiment, the first lateral distance di at,i can be determined from data associated with the first detected object OBJ1 and / or measured on the first acquired image, such as a size (in pixels) of the first detected object OBJ1 and / or an angle between a pixel belonging to the first detected object OBJ1 and a predefined reference pixel on the first acquired image. Such data can be estimated by a classifier having detected the first object OBJ1. Such data can also be estimated from measurements made on the first acquired image. For example, the first lateral distance di at,i can be determined from the estimation of a first position associated with the detected object OBJ1. Such a first position can for example be a position of a first reference pixel b1 belonging to the detected object OBJ1. Such a first reference pixel b1 is for example represented in Figure 5. In one embodiment, such a first reference pixel b1 can be identified by the object classifier as being the wheel of the vehicle for example. Alternatively, the first reference pixel b1 can be identified by measuring a minimum angle 0min, as represented in Figure 11. To measure such a minimum angle 0min, a vanishing point C' of the first acquired image is considered, such a vanishing point coinciding with the optical center of the first acquired image in Figure 11, as well as a vertical axis VL passing through such a vanishing point C'.The first reference pixel b1 can then be selected from among the set of pixels forming the detected object OBJ1, so that the angle formed between, on the one hand, the vertical axis VL and, on the other hand, the axis formed between the vanishing point C' and such a reference pixel b1, is minimal 0min.

[0096] The first reference pixel b1 thus obtained then has a first position in the first acquired image, which can be noted by the coordinates of the point b1 in the image frame (y,z). From such coordinates in the image frame (y,z), the first position associated with the detected object OBJ1 can be estimated in the world frame (X,Y,Z). In particular, the coordinates of the first position in the world frame (X,Y,Z) can be determined, from the coordinates of the first reference pixel b1 in the image frame (y,z) and the predefined characteristics of the camera 1, on the basis of the pinhole model. The application of the pinhole model so as to obtain the coordinates associated with the first detected object OBJ1 in the world frame (X,Y,Z) is based in particular on several calculation hypotheses linked to the properties of the camera 1 as well as to the ENV environment associated with the world frame (X,Y,Z). For this, classic extrinsic parameters of the camera 1 can be considered. A fish-eye type camera 1 can be used. In addition, a flat world assumption can be considered. Alternatively, georeferencing of the ENV environment by ground marking or other topographic surveys can also be considered.In the context of the applied pinhole model, the geometric distortions possibly induced by the optical system of the camera 1 can be neglected. In another embodiment, the determination of the coordinates associated with the first detected object OBJ1 in the world frame (X,Y,Z) can comprise a correction of the distortion of the lens of the camera 1. The first lateral distance di. a t,i associated with the first detected object OBJ1 (real, expressed in meters) can then be directly deduced from the coordinate along the Y axis of the first position according to the world reference frame (X,Y,Z). The first lateral distance di at,i associated with the first detected object OBJ1 can in particular be obtained by calculating the difference between the coordinate along the Y axis of the first position according to the world reference frame (X,Y,Z) and the coordinate along the Y axis of a point of origin of the camera 1 of the main vehicle VP, for example coincident with the vanishing point C', the optical center of the image, or even a point on the image associated with the location of the camera 1 in the environment ENV. Alternatively, a combination of the aforementioned methods can be implemented in step 420 in order to determine the first lateral distance di at ,i associated with the first detected object OBJ1.

[0097] At the end of step 420, the first detected object OBJ 1 on the first image acquired at the first acquisition time T1 can then be associated with data relating to a first position of the first object OBJ1. In particular, such data can be a first lateral distance di at,i. Such a first lateral distance di at ,i corresponds in particular to an estimation of a lateral difference separating the main vehicle VP from the detected object OBJ1. In other words, step 420 makes it possible to estimate a real lateral difference (i.e. in the environment ENV) between the main vehicle VP and a neighboring moving body corresponding to the first detected object OBJ1 (such a neighboring body being for example a neighboring vehicle VV1 of the main vehicle VP). In the remainder of the description, it is considered that such a neighboring moving body corresponding to the first detected object OBJ1 is a neighboring vehicle VV1.

[0098] In a step 430, a set of parameters associated with the neighboring vehicle VV1 thus detected can be determined and / or estimated by the device 2. Such parameters can in particular be estimated from the first lateral distance di at,i associated with the first detected object OBJ1 and corresponding to the neighboring vehicle VV1. Such parameters may for example include a current position p n associated with the neighboring vehicle VV1 detected, a current speed v n associated with the neighboring vehicle VV1 detected and / or current dimensions s n associated with the neighboring vehicle VV1.

[0099] Thus, in step 430, determining a set of parameters associated with the neighboring vehicle VV1 may include determining a current speed v n associated with the neighboring vehicle VV1. In one embodiment, such a current speed v n can for example be a rotational speed estimated from the observation of the evolution of a point located on the wheel by the wheel classifier. Alternatively, such a current speed v n can for example be deduced from the first lateral distance di at,i associated with the first detected object OBJ1 as determined in step 420, as well as a temporal data item iTTC associated with the first detected object OBJ1. Such a temporal data item iTTC may relate to an inverse magnitude of a collision time associated with the first detected object OBJ1. In the case of a first detected object OBJ1 corresponding to a segment (i.e., to a set of optical flows), such a temporal data item iTTC associated with each segment may for example be determined by one of the methods described in documents WO 2018059629, WO 2018059631 and WO 2018059632.

[0100] Thus, at step 430, a current speed v n associated with the neighboring vehicle VV1 can be determined as a speed associated with the first detected object OBJ1: v n = d iat:1 x iTTC where: - v nis a relative speed (with respect to the main vehicle VP) associated with the first detected object OBJ1 (expressed in meters per second), - d iat is the lateral distance associated with the first detected object OBJ1 (expressed in meters), and - iTTC is the time data associated with the first detected object OBJ1 (expressed as the inverse of a time).

[0101] Additionally, in step 430, determining a set of parameters associated with the neighboring vehicle VV1 may include determining current dimensions s n associated with the neighboring vehicle VV1. In one embodiment, such current dimensions n can be predefined by pre-existing dimensions typically observed in vehicles. Such pre-existing dimensions may, for example, correspond to a predefined length L1, a predefined width L2 and a predefined height H.

[0102] Finally, in step 430, determining a set of parameters associated with the neighboring vehicle VV1 may include determining a current position p n associated with the neighboring vehicle VV1. In particular, such a current position p n can be associated with an estimated point on the neighboring vehicle VV1, for example a first reference point Wi associated with the neighboring vehicle VV1. Such a first reference point Wi can in particular be associated with the center of a front face (not visible on the acquired image of figure 5) of the neighboring vehicle VV1. Thus, the estimation of the current position p n associated with the neighboring vehicle VV1 can return to estimating the coordinates (X_Wi,Y_Wi,Z_Wi) of the first reference point Wi in the world frame (X,Y,Z). In other words, in one embodiment: P„(X) = X_W, Pn (Y) = Y_Wi , Pn (Z) = Z_W 1

[0103] To estimate the coordinates (X_Wi,Y_Wi,Z_Wi) of the first reference point Wi, in an embodiment represented by FIG. 12, the device 2 can rely on the coordinates (X_C, Y_C) of a central point C of the camera 1 in the world frame (X,Y,Z), such a central point C corresponding for example to the optical center or to a point associated with the location of the camera 1 or of a sensor of the camera 1 in the environment ENV (the projection of such a central point C resulting in the point C' in FIG. 5). The device 2 can also rely on the pre-existing dimensions L1, L, H of the neighboring vehicle VV1, on the first lateral distance determined in step 420, on the focal length f of camera 1 and on the size of the sensor sz_capt in particular. Device 2 can then determine, by applying the pinhole model as well as Thales' theorem: Or : - (X_W 1 , Y_W 1) are the coordinates of the first reference point Wi in the world frame (X,Y,Z), - (X_C, Y_C) are the coordinates of the central point C of camera 1 in the world frame (X,Y,Z), - di atil is the first lateral distance associated with the first detected object OBJ1, - f is the component along the X axis of the focal length of camera 1, - sz_capt is the actual size of the sensor, and - L2 is the predefined width of the neighboring vehicle VV1.

[0104] Such coordinates in the world frame (X,Y,Z) of the first reference point Wi can in particular be determined by considering that the central point C is coincident with the center of the acquired image. Optical deformations can be considered neglected. Furthermore, the focal length f is considered identical on both axes - vertical VL and horizontal - of the image. In particular, the aforementioned formulas describing the coordinates of the first reference point Wi depend on the relative positioning of the neighboring vehicle VV1 with respect to the main vehicle VP.

[0105] Finally, a Z_W coordinate of the first reference point Wi along the Z axis in the world frame (X,Y,Z) can be determined as corresponding to half the predefined height of the neighboring vehicle VV1, namely Z_W = H / 2. Alternatively, the Z_W coordinate can correspond to the coordinate along the Z axis of the first position associated with the first detected object OBJ1. The Z_W coordinate can in particular correspond to the height in the world frame (X,Y,Z) of the point closest to the ground whose projection in the image frame (y,z) belongs to the first detected object OBJ1.

[0106] Thus, at the end of step 430, the device 2 can obtain a set of (current) parameters (p n , v n , s n ) associated with the neighboring vehicle VV1 (and therefore with the first detected object OBJ1). In particular the parameters (p n , v n , s n) are associated with a given acquisition time - namely the first acquisition time T1 here -, corresponding to a certain positioning and certain kinematic conditions of the neighboring vehicle VV1 in the environment ENV, estimated from the first detected object OBJ1 at such a first acquisition time T 1 .

[0107] Optionally, at the end of step 430, a step (not shown in FIG. 4) may include verifying the plausibility of the detection of a neighboring vehicle VV1 from the detected object OBJ1 as well as the parameters (p n , v n , s n ) determined from the first detected object OBJ1. A plausibility test can then be optionally implemented, so as to, for example: - check that the current speed v n reflects a neighboring body moving in a plausibly similar manner to a neighboring vehicle VV1 and / or in a direction close to that of the main vehicle VP, - verify that a height of the set of pixels forming the first detected object OBJ1 has a plausible height to belong to a neighboring vehicle VV1 (for example, a first detected object OBJ1 corresponding to the movement of a bird captured in the field of view FOV will potentially have a height not belonging to an interval of plausible height values ​​determined as corresponding to a vehicle).

[0108] In a step 440, the device 2 can check whether the neighboring vehicle VV1 and / or the first detected object OBJ1 corresponds to a vehicle already previously detected by the device 2, so as to associate the first detected object with an object already tracked. Generally, at any given acquisition time, the parameters associated with a detected vehicle, as determined in step 430, allow the device 2 to estimate an occupation space of the environment ENV by the detected vehicle. For example, such an occupation space of the detected vehicle can be estimated from the position of the vehicle (for example estimated by a conventional classifier or determined in step 430) as well as the dimensions associated with this vehicle (for example the pre-existing dimensions). Such an occupation space can for example correspond to a set of coordinates according to the world reference frame (X,Y,Z) in the environment ENV associated with the detected vehicle.For example, with reference to Figure 2, the neighboring vehicle VV2 is already detected by the device 2 at the first acquisition time T1, for example by a conventional classifier detecting and following the front faces of the vehicles. An occupation space corresponding to a set of positions in the environment ENV is then already stored by the device 2 at the stage of the first acquisition time T1, in association with such a neighboring vehicle VV2. Thus, at the first acquisition time T1, a position associated with an object detected at the first acquisition time T1, for example the first position associated with the first detected object OBJ1 at step 410, can then be compared to the occupation space associated with the neighboring vehicle VV2, so as to verify whether the object detected at the first acquisition time T1 corresponds to the movement of the neighboring vehicle VV2.

[0109] Thus, in step 440, a verification of a detected object can for example be implemented by comparing the first position associated with the first detected object OBJ1 and / or the set of parameters associated with the neighboring vehicle VV1 estimated from the first detected object OBJ1 with one or more occupation spaces (i.e., one or more sets of coordinates) respectively associated with one or more bodies (e.g., vehicles) previously detected and stored by the device 2.

[0110] If, in step 440, such a first position associated with the first detected object OBJ1 and / or the current position of the neighboring vehicle VV1 is included in an occupation space stored by the device 2, then the device 2 determines that the first detected object OBJ1 (and therefore the neighboring vehicle VV1) corresponds to a known object of interest (e.g., a vehicle already tracked) of the device, which has already been detected at a time preceding the first acquisition time T1. The device 2 then does not need to create a new virtual object associated with the first detected object OBJ1 and can rely on the data already associated with the known object of interest. Steps 400 to 440 then make it possible to carry out tracking of such an object of interest (for example, the neighboring vehicle VV1) associated with the first detected object OBJ1. Tracking an object of interest includes, in particular, estimating future parameters associated with such an object of interest (for example, a future position, future dimensions, a future speed, etc.) and updating the object of interest. For this, in a step 450, the parameters (p n , v n , s n ) associated with the neighboring vehicle VV1 and the first acquisition time T1 as determined from the first detected object OBJ1 can be fed to a Kalman filter (for example an Extended Kalman Filter or EKF). Such a Kalman filter is in particular configured to estimate future (or predicted) parameters (p n+1 , v n+1 , s n+1 ) of an object of interest from at least current parameters (p n , v n , s n ) provided by device 2.

[0111] Once an object of interest is detected and tracked, the Kalman filter or other data estimator may estimate future parameters associated with the object of interest. For example, at the end of step 450, feeding the parameters (p n , v n , s n) associated with the neighboring vehicle VV1 to the Kalman filter allows tracking of the neighboring vehicle VV1 on the basis of successive predictions of the Kalman filter as to a position and a speed of the neighboring vehicle VV1 in the environment ENV. In particular, a lateral object distance (corresponding to a lateral gap evolving between the main vehicle VP and the neighboring vehicle VV1 tracked) can be predicted. Thus, any object of interest tracked by the device 2 can be associated, depending on the instant considered, with a given lateral object distance, as well as with a set of predicted parameters making it possible to estimate a position and a movement of the object of interest tracked relative to the main vehicle VP, as the object of interest moves in the environment ENV.

[0112] At a step 470, from the estimated future parameters for a tracked vehicle, the device can then proceed to update the parameters associated with the tracked vehicle corresponding to the occupied space. Such a step 470 of updating the parameters associated with a vehicle already detected will be described in detail in the remainder of the description.

[0113] If, in step 440, the first position associated with the first detected object OBJ1 and / or the current position of the neighboring vehicle VV1 is not included in an occupation space stored by the device 2, then the device 2 can proceed to a step 460 of creating a virtual object OV1 associated with the first detected object OBJ1 (and therefore with the detected neighboring vehicle VV1). An example of a virtual object OV1 created from the first detected object OBJ1 is illustrated in FIG. 6. In particular, the virtual object OV1 can be created from the parameters (p n , v n s n) determined in step 430. For example, from the determined coordinates (X_Wi, Y_Wi, Z_Wi) of the first reference point Wi and using the model of the camera 1, a first virtual reference point wi having coordinates (ywi.zwi) in the image frame (y,z) can be placed on the acquired image, as shown in the Figure 6. In particular, the virtual point wi can be positioned in the image frame (y,z) under several hypotheses: - the first detected object OBJ1 corresponds to a rear portion of the neighboring vehicle VV1 (eg, it includes a rear wheel of the neighboring vehicle VV1), - the virtual reference point wi is considered positioned on a lateral edge of the acquired image. In other words, the lateral portion visible on the acquired image is considered to represent the entire length of the neighboring vehicle VV1 in the state considered on the acquired image and the first reference point Wi is considered positioned on the front face of the neighboring vehicle VV1.

[0114] In another embodiment, the first virtual reference point wi may be placed elsewhere than on the side border of the image, for example inside the image or outside the image.

[0115] Thus, with reference to Figure 6, if the pixel located at the bottom left of the acquired image is considered to be the origin of the image reference frame (y,z) and the neighboring vehicle VV1 is detected as being on the right lateral side of the main vehicle VP (as illustrated in Figures 2 and 5 to 12), we have ywi = 0, as illustrated in Figure 6.

[0116] In addition, the zwi coordinate of the first virtual reference point wi along the z axis can be determined from the z-axis coordinate of the first position to determine the lateral distance di at l , in the image frame. In other words, the virtual reference point wi may have the same height along the z axis as the detected object OBJ1. In one embodiment, the corresponding height zwi of the virtual reference point wi along the z axis may be predefined and stored by the device 2 as corresponding to a predefined ground height. In one embodiment, the height zwi may also differ from the height of the detected object OBJ1.

[0117] From such a virtual reference point wi positioned on the acquired image and the current dimensions s nassociated with the vehicle (corresponding here to the pre-existing dimensions L1, L2, H) associated with the neighboring vehicle VV1, the virtual object OV1 can be created, having coordinates (ywi.zwi) in the two-dimensional space (y,z). In particular, the virtual object OV1 can substantially include the first detected object OBJ1, since the latter is identified as corresponding, at least partially, to the neighboring vehicle VV1 detected on the acquired image. Such a virtual object OV1 is for example represented in Figure 6. The virtual object OV1 created then corresponding to a modeling, on the acquired image (therefore in the image frame (y,z)), of the occupation space in the first acquired image of the neighboring vehicle VV1 detected from the first detected object OBJ1. The current position p n of the neighboring vehicle VV1 can, equivalently, refer to the position of the reference point Wi or to the virtual reference position wi.

[0118] Optionally, the coordinates (ywi.zwi) of the virtual object OV1 thus created can be fed to the Kalman filter, in association with the parameters (p n , v n , s n ), the detected neighboring vehicle VV1 and the first acquisition time T1. Thus, predicted parameters associated with the created virtual object OV1 can be estimated as the object of interest (typically, the neighboring vehicle VV1) corresponding to the virtual object OV1 moves in the environment ENV. In particular, an object lateral distance di at 0V1 can be associated with the virtual object OV1. For example, at the first acquisition time T1 having allowed the creation of the virtual object OV1 from the first detected object OBJ1, the lateral object distance di at 0V1 can correspond to the first lateral distance di at l . The value of the lateral distance object di at 0V1can then evolve according to the successive predictions of the data estimator estimating the movement of the neighboring vehicle VV1 followed (and modeled by the virtual object OV1).

[0119] Optionally, such a virtual object OV1 associated with a detected neighboring vehicle VV1 can then be displayed on the communication interface 50 corresponding to a display screen, for example by superimposing the virtual object OV1 on a display of the image stream coming from the camera 1, as illustrated in FIG. 6. Thus, a driver of the main vehicle VP can quickly identify the occupation space of the neighboring vehicle VV1 detected in his field of view FOV. In particular, like the first position associated with the detected object OBJ1, the lateral distance associated with the detected object OBJ1 or even parameters (p n , v n , s n) associated with the neighboring vehicle VV1, such a virtual object OV1, as represented in figure 6, is associated with a given acquisition time - here the first acquisition time T1 - linked to a certain positioning and to certain kinematic conditions of the neighboring vehicle VV1 in the environment ENV, estimated from the detected object OBJ1 at such a first acquisition time T1.

[0120] The anticipated tracking method of FIG. 4 is now considered at a second acquisition time T2, subsequent to the first acquisition time T1. For example, a first implementation of the aforementioned steps having led to the creation of a virtual object OV1 associated with the first acquisition time T1 and to a first detected object OBJ1, a second acquisition time T2 is considered. At the second acquisition time T2, a second object OBJ2 can be detected at step 410. Such a second object OBJ2 can for example be detected on a second image acquired at a step 400 during the second acquisition time T2, the second image being distinct from the first image on which the first object OBJ1 was detected. The detection of such a second object OBJ2 on a second image is for example represented in figure 8. The first object OBJ 1 associated with the first acquisition time T1 is also represented there for comparison.

[0121] As previously described, a second lateral distance di at 2 associated with the second detected object OBJ2 can be determined in step 420, a second reference point W2 and parameters (p n *, v n *, s n *) associated with the second detected object OBJ2 can be determined in step 430.

[0122] In step 440, the object verification is implemented on the second detected object OBJ2, so as to identify whether the second detected object OBJ2 can be associated with an existing (already created) virtual object. In particular, in step 440, it can be determined that the second detected object OBJ2 at the second acquisition time T2 can be associated, by virtue of its parameters (p n *, v n *, s n*) determined in step 430, to the movement of the neighboring vehicle VV1 already previously detected by the device 2 (namely at the first acquisition time T1), and represented by the virtual object OV1, as illustrated in figures 6, 7 and 9.

[0123] At step 450, the parameters (p n *, v n *,s n *) associated with the second detected object OBJ2 can also be fed to the Kalman filter. Thus, the predicted parameters (p n+ i, v n+1 , s n+1 ) by the Kalman filter for the virtual object OV1 can notably take into account the parameters (p n *, v n *, s n *) associated with the second detected object OBJ2.

[0124] Step 470 of updating the virtual object OV1 associated with the neighboring vehicle VV1 at the second acquisition time T2 is now detailed. Such a step 470 of updating the virtual object OV1 makes it possible in particular, at the end of step 470, to obtain a set of parameters (p n +i*,v n +i*,Sn+i*) allowing to characterize the position and the movement of the virtual object OV1 (and therefore of the neighboring vehicle VV1) in the environment ENV at the second acquisition time T2. The virtual object OV1 to be updated is associated with current parameters (p n , v n , s n ) to be updated.

[0125] Step 470 of updating the virtual object OV1 takes into account in particular the predicted (or future) state and parameters (p n +i, v n +i, s n +i) of the virtual object OV1 provided by the data estimator. Such predicted parameters (p n +i, v n +i, s n+i) reflect an estimate, for example provided by the Kalman filter of the device 2, of the state of the neighboring vehicle VV1 as predicted from the data from the first acquisition time T1. In particular, the predicted state of the virtual object OV1 may also include a lateral object distance di at ,predicted ovi, corresponding to a lateral deviation estimated via the Kalman filter between the neighboring vehicle VV1 and the main vehicle VP at the second acquisition time T2.

[0126] In the context of the proposed method, step 470 of updating the virtual object 407 comprises sub-steps 471, 472 of updating a length of the virtual object OV1 according to a first predefined criterion and sub-steps 473, 474 of updating a position of the virtual object OV1 according to a second predefined criterion.

[0127] At a sub-step 471, the predicted, or future, parameters (p n +i, v n +i, s n+i) of the virtual object OV1 as estimated by the Kalman filter are compared to the current parameters (p n , v n , s n ) of the virtual object OV1. In particular, the current position p n associated with the virtual object OV1 and the future position p n +i estimated for the virtual object OV1 are compared. If, in substep 471, a first criterion is satisfied following such a comparison between the current position p n and the future position p n +i, an update 472 of the current dimensions s n , and in particular of a current length s n (X), of the virtual object OV1 is implemented. In particular, such a first criterion is satisfied if the future position p n +i predicted for virtual object OV1 indicates that virtual object OV1 has moved back in the environment ENV relative to its current position p n. For example, the virtual object OV1 can be considered as having a position that is moved back from its current position p n if the lateral component of its predicted position p n +i(X) is less than the lateral component of its current position p n (X) in the (X,Y,Z) frame. In such a case, the device 2 maintains the state of the virtual object OV1 except for its current dimensions s n by implementing a systematic extension of the length of the virtual object OV1 associated with the neighboring vehicle VV1.

[0128] In substep 472, updating a current length s n of the virtual object OV1 is then implemented. The current dimensions n of the virtual object OV1 associated with the neighboring vehicle VV1 are then updated by modifying a length associated with the neighboring vehicle VV1, such a length being for example the coordinate along X of the dimensions n ■ sn+l( )* = S n (X) + 8 with: 8 = \p n+1 (X) - p n (X)\ where: - 8 is a difference in length (expressed in meters), - s n+1 (X)* is the updated length of the neighboring vehicle VV1 in the world frame, - s„(X) is the current length of the neighboring vehicle VV1 in the world frame, - p n+1 (X) is the coordinate along the X axis of the position of the reference point predicted W2 by the Kalman filter in the world frame, - p n X) is the X-axis coordinate of the position of the first reference point W1 in the world frame.

[0129] In particular, the length s n (X) corresponds to a longitudinal component (along the X axis) of the dimensions associated with the virtual object OV1.

[0130] In other words, sub-step step 472 comprises an extension of the length of the virtual object OV1 modeling the neighboring vehicle VV1. The difference in position of the reference point W associated with a front face of the neighboring vehicle VV1 between a current position p n (X and a predicted position p n+1 ( ) is then reported in length 8.

[0131] In particular, such an extension at step 472 is implemented with: W 2 * = WL where: - w 2 * is the updated reference point associated with the updated virtual object OV*, - w ± is the first reference point associated with the virtual object OV1 before update.

[0132] In other words, the updating of the length of the neighboring vehicle VV1 in step 472 is implemented while maintaining the position of a front face of the virtual object OV*. Thus, the updating of the virtual object OV1 in OV* makes it possible to update a length of the neighboring vehicle VV1 by preventing the virtual object OV1 from “moving back” in the acquired image.

[0133] In particular, updating the dimensions s n of the virtual object at substep 472 maintains the current speed v n constant.

[0134] So, the updated settings (p n+1 * , v n+1 *, s n+1 *) at the end of sub-step 472 satisfy:

[0135] Such a length update s n (X) of the virtual object OV1 associated with the neighboring vehicle VV1 in substep 472 is represented in figure 7.

[0136] If, in sub-step 471, the first criterion is not satisfied, no update of the length is implemented and step 470 of updating the virtual object OV1 continues with sub-steps 473, 474.

[0137] At a substep 473, an update of the position p n of the virtual object OV1 is implemented if a second criterion is satisfied. Such a second criterion depends in particular on a second comparison between the second lateral distance di a t,2 determined from the second detected object OBJ2 and the lateral object distance di at ,ovi estimated for the virtual object OV1. In particular, if, in substep 473, the second lateral distance di a t,2 is strictly less than the lateral object distance di at ,ovi, a substep of updating 474 the current position p nof the virtual object OV1 is implemented, the virtual object OV1 being repositioned so that its lateral distance object di at ,ovi* up to date is minimal (ie, di at ,ovi* = min(di at ,ovi, di at ,2))- Indeed, such an update 474 allows to rectify a lateral distance object di at ,ovi between the neighboring vehicle VV1 and the main vehicle VP having been overestimated during the previous iteration and an update of the position of the virtual object OV1 is required. The update of the current position p n of the neighboring vehicle VV1 then includes in particular at least one update of a lateral component p n (Y) (i.e. along the Y axis in the world frame (X,Y,Z)) of the current position p n of the neighboring vehicle VV1. In one embodiment, updating the current position p n may also include an update of other components of the current position p n, for example a height component p n (Z).

[0138] In a substep 474, if the second criterion is satisfied, updating such an updated position includes associating an updated reference point w 2 * associated with the updated virtual object OV*, such an updated reference point w 2 * being different from the first reference point Determining the updated reference point w 2 * then rests on the second lateral distance di at 2 . The positioning of such an updated reference point w 2 * in particular respect the same constraints on the image as for the first reference point w ± .

[0139] In particular, updating the parameters in substep 474 maintains a constant image stream.

[0140] So, the updated settings (p n +i* , v n+1 *, s n+1*) at the end of sub-step 474 can satisfy:

[0141] In particular, the expression of the up-to-date parameters (p n+1 *, v n+1 *,s n+1 *) at the end of sub-step 474 depends on the definition of position p n and reference points wi, W2*. For example, if point p n corresponds to the position of a first reference point wi positioned in the middle of the virtual object OV1, then the updated parameters (p n+1 *, v n+1 *, s n+1 *) at the end of sub-step 474 can satisfy the following relationship:

[0142] In one embodiment, at substep 474, the size s n+1 * can also be updated, for example to maintain a constant occupation space of the updated virtual object OV* in the image:

[0143] Such an update 474 of the position of the virtual object OV1 so as to obtain an up-to-date virtual object OV* is illustrated in Figure 9, for which the second is satisfied (the second lateral distance di a t,2 is less than the lateral object distance di at ,ovi.

[0144] So such an update 474 changes the position p n , the lateral distance object di at ,ovi and thus the speed v n of the virtual object OV1, while keeping a parameter of the virtual object OV1 constant, for example the collision time, the ITTC, or p n (X) * v n .

[0145] If, in substep 473, the lateral object distance di at ,ovi is less than the second lateral distance di at ,2, the update 474 of the position of the virtual object OV1 and an update of the virtual object OV1 directly on the basis of the predictions (p n +i, v n +i, s n+i) estimated by the Kalman filter can be implemented at a classical update step 475 of the virtual object OV1.

[0146] Thus, the anticipated tracking method described in Figure 4 allows anticipated tracking of the neighboring vehicle before a front face of the neighboring vehicle VV1 is visible in the field of view FOV. In particular, the proposed method allows a specific update of the length and / or position of the virtual object OV1 modeling the neighboring vehicle VV1 so as to compensate for and reduce overestimated values ​​of lateral distance separating the neighboring vehicle VV1 and the main vehicle VP, which can lead to dangerous situations (for example, an overestimation of the lateral distance positions the neighboring vehicle VV1 further from the main vehicle than it actually is). Furthermore, the update of a length of the virtual object OV1 can be implemented even in the absence of new detection (for example, of a second detected object OBJ2).

[0147] Optionally, the anticipated tracking method as proposed can be implemented in combination with tracking implemented by a classifier. In other words, the anticipated tracking method described in Figure 4 can be implemented even when a front face of the neighboring vehicle VV1 becomes visible, for example at the third acquisition time T3 in Figure 10. Alternatively, the anticipated tracking method can be implemented as long as at least reliable tracking of the neighboring vehicle VV1 is not yet possible, for example as long as a front face of the neighboring vehicle VV1 is not yet visible (for example, in the scenarios of the first and second acquisition times T1, T2).For this, the parameters fed to the Kalman filter in step 440 can be associated with a variance value chosen so that when parameters estimated by a frontal face classifier are fed to the Kalman filter, the parameters estimated via the proposed feedforward tracking method are no longer significant for the filter to predict the parameters. The method of FIG. 4 can therefore be iterated as long as images are acquired by the device 2, or as long as the parameters fed to the Kalman filter are used to predict the parameters (for example, as long as data from classifiers, having a lower uncertainty value and therefore a lower variance or covariance for example, are not yet fed to the Kalman filter).

Claims

Claims

1. Method for tracking at least one neighboring vehicle (VV1) present in an environment (ENV) of a main vehicle (VP), said neighboring vehicle (VV1) and said main vehicle (VP) being motor vehicles, the method being implemented by a device (2) configured to provide a driving assistance function for the main vehicle (VP), said device (2) being connected to at least one camera (1) on board the main vehicle (VP) and capable of acquiring images of a scene surrounding the main vehicle (VP) according to at least one field of view (FOV) and at acquisition times (T 1 , T2, T3), the method comprising the following steps: - detecting (410), from at least a first image acquired at a first acquisition time (T1), a first object (OBJ1) having a vertical plane extending laterally relative to the main vehicle (VP), - determine, according to a predefined coordinate system (X,Y,Z), at least one piece of data relating to a first position associated with the first detected object (OBJ1), - create (460) a virtual object (OV1) associated with a set of parameters (pn,vn,sn) comprising at least one current position (pn) associated with the virtual object (OV1), said current position (pn) being estimated (430) from at least said data relating to the first position according to the predefined coordinate system, - estimate (440), from at least said current position of the virtual object (pn), a future position (pn+1) of the virtual object (OV1), and - updating (470, (pn+1*,vn+1*,sn+1*)) the parameters associated with the virtual object (OV1) from at least said future position (pn+1), said updating (470) of the parameters associated with the virtual object (OV1) comprising, if a first criterion is satisfied (471), an updating (472, sn+1*) of a current length (sn(X)) of the virtual object (OV1).

2. Method according to claim 1, wherein updating the current length of the virtual object (OV1) comprises extending the length of the virtual object (OV1) by a value determined from a gap (5) between the current position (pn) and the future position (pn+1), while maintaining the current position (pn) of the virtual object (OV1) constant.

3. Method according to any one of the preceding claims, wherein the updating of the current length of the virtual object (OV1) is implemented if a first comparison between the future position (pn+1) and the current position (pn) indicates a receding position of the virtual object (OV1) in the first acquired image relative to the current position (pn).

4. Method according to any one of the preceding claims, wherein said data relating to the first position comprises a first lateral distance (dlat, 1) associated with the first detected object (OBJ1) for the first acquisition time (T1), said first lateral distance (dlat, 1) being determined from measurements of the first detected object (OBJ1) in the first acquired image.

5. Method according to any one of the preceding claims, in which the future position (pn+1) is associated with an object lateral distance (dlat,OV1) of the virtual object (OV1).

6. A method according to the preceding claim, further comprising: - obtaining, from at least one second image acquired at a second acquisition time (T2) subsequent to the first acquisition time (T1), data relating to a second position associated with a second detected object (OBJ2), said data relating to the second position comprising a second lateral distance (dlat, 2) associated with the second detected object (OBJ2) for the second acquisition time (T2), - if the second detected object (OBJ2) is determined to belong to the virtual object (OV1), comparing, in a second comparison, the second lateral distance (dlat, 2) and the object lateral distance (dlat,OV1), wherein the update (470, (pn+1*,vn+1*,sn+1*)) of the parameters associated with the virtual object (OV1) further depends on a result of the second comparison.

7. Method according to claim 6, in which, if the second lateral distance (dlat, 2) is strictly less than the object lateral distance (dlat, OV1), the update (470, (pn+1*, vn+1*, sn+1*)) of the parameters associated with the virtual object (OV1) comprises an update (474, pn+1*) of the current position (pn) of the virtual object (OV1), said update (474, pn+1*) of the current position (pn) including at least one update of a lateral component (pn(Y)) of the current position (pn) of the virtual object (OV1) corresponding to a repositioning of the virtual object (OV1).

8. A method according to any preceding claim, wherein the current position (pn) of the virtual object (OV1) is associated with the position of a first reference point (w1) belonging to the virtual object, said position of a first reference point (w1) belonging to the virtual object (OV1) being determined such that, in the predefined coordinate system corresponding to an image reference (x,y) in the acquired images, the first reference point (w1) is positioned on a vertical end border of the first image.

9. Method according to any one of the preceding claims, in which the parameters (pn,vn,sn) associated with the virtual object (OV1) are updated (470, (pn+1*,vn+1*,sn+1*)) while maintaining a parameter of the virtual object (OV1) constant (p n * v n ).

10. Device (2) configured to provide a driving assistance function for a main vehicle (VP), said device (2) being connected to at least one camera (1) on board the main vehicle (VP) and capable of acquiring images of a scene surrounding the main vehicle (VP) according to at least one field of view (FOV) and at acquisition times (T1, T2, T3), in which the device (2) comprises at least one processing circuit (20, 30, 40, 50, 60) configured to implement a method for tracking at least one neighboring vehicle (VV1) present in an environment (ENV) of the main vehicle (VP) according to one of claims 1 to 9.

11. Computer program comprising instructions for implementing the method according to one of claims 1 to 9 when this program is executed by at least one processor (21, 31, 41, 51, 61).

Images